Intelligence and personas
| Term | Definition |
|---|---|
| Intelligence | The home chat surface where you ask questions, continue threads, attach files, mention personas or knowledge bases with @, and use composer presets such as Browse, Agentic, and Civilization. |
| Persona | A custom AI identity: name, role, system prompt, model choices, tool permissions, and optional knowledge base. Switching personas changes behavior without reloading the app. See Personas. |
| System prompt | Instructions that set a persona’s tone, role, and constraints — for example, “you are a research assistant that cites sources.” |
| Resident persona | A persona stored and run on your device. Your prompts and context stay local for that persona. Includes the default General Assistant and every persona you create yourself. |
| Connected persona | A persona hosted on another device — typically after pairing with a contact. You send messages from your device; their device runs inference. The host’s system prompt is not sent over the network. The host decides whether the persona can be invoked remotely and whether a portable copy is offered. |
| Portable persona | A downloaded copy of a shared persona definition on your device: name, description, system prompt, model references, and tool policy. Knowledge base documents are not included — only the persona definition travels. To use the host’s documents, keep using the persona connected on their device, or attach your own knowledge base locally. |
| Knowledge base | A collection of documents you ingest to ground persona answers. Bind a knowledge base to a persona for citations and retrieval in Intelligence chat. |
| Single | One persona responds per turn when multiple personas are tagged. You choose Reply as to pick which participant speaks. |
| Parallel | Every tagged persona answers at the same time. Replies appear in a side-by-side panel so you can compare perspectives. |
| Review | A two-step pipeline: the first tagged persona drafts, the second reviews and refines (with optional additional review rounds). With three or more personas, you can choose which two act as drafter and reviewer. |
| Agentic mode | A composer preset where the AI can plan, use tools, and take multiple steps before answering, with visible progress. |
| Agentic skill | A template that shapes how an agentic run behaves — bundled skills for writing, analysis, and more, or custom skill files you author. |
| On-device | Inference runs on your machine. Intelligence shows an on-device indicator when the reply is generated locally. |
Local AI and models
| Term | Definition |
|---|---|
| Model tier | A grouping of AI models by size and memory requirements. The system recommends a tier based on your device’s capabilities. See On-Device AI. |
| Token | The basic unit of text an AI model processes. Words are broken into one or more tokens. Token counts and tokens-per-second appear in the UI as performance indicators. |
| LoRA (Low-Rank Adaptation) | A small adapter file loaded on top of a base model to specialize behavior without downloading a full new model. Personas can optionally attach a LoRA adapter. See On-Device AI. |
| MLX | Apple’s on-device inference backend, used on Apple Silicon Macs when available. |
| GGUF | A compact model file format used with the llama.cpp inference backend on supported desktops. |
| WebGPU | A GPU inference path used by the Oasis local AI runtime when a compatible GPU is available. |
| Oasis | The local AI runtime that powers on-device chat, personas, and agent capabilities in the desktop app. |
Identity and contacts
| Term | Definition |
|---|---|
| ODID (On-Device ID) | Your unique device identifier. Generated locally and used to identify you to contacts and in spaces. No account or sign-up required. |
| Pairing | Adding a contact securely by exchanging cryptographic keys — for example, with a 6-digit rotating code. Pairing can also exchange capability cards so connected and portable personas from the other device appear in your pickers. |
| Contact | Someone on your contacts list after pairing or adding by ODID. You can message them, see presence, and use their shared personas when they allow it. |
Collaboration and networking
| Term | Definition |
|---|---|
| Peer-to-peer (P2P) | Devices connect directly for collaboration features such as spaces, direct messages, file sharing, and voice. Your live session syncs between participants rather than living on a central collaboration server. Some discovery features (finding a room or a device on your network) may use directory or LAN services, but the active session stays between the devices in the room. |
| Space | A shared realtime workspace. Host or join with a code, work in shared apps (Documents, Whiteboard, and more), chat, use voice, and share files with everyone in the room. |
| Host | The participant who creates a space and shares the join code. If the host disconnects, the space can keep running when backup relay is enabled. |
| WebRTC | The technology behind realtime voice and live media between devices in spaces and direct sessions. |
| CRDT (Conflict-Free Replicated Data Type) | The data structure that keeps shared app state in sync when several people edit at once. Changes from different participants merge automatically without manual conflict resolution. |
| Edit lock | A short-lived lock (up to 30 seconds) on a specific field or element while someone is editing it, so two people do not change the same field at the same time. Locks release automatically when the editor moves on. |
| Local discovery | Finding nearby spaces or contacts on the same local network (via mDNS) or, on supported devices, within Bluetooth range — without needing a join code. |
| End-to-end encryption (E2E) | Encryption where only the sender and recipient can read the content. Direct messages and private file transfers use E2E. See About webAI. |
| Relay (backup host) | A peer that can help keep a space connected if the original host disconnects. Enable under Sharing & Effects. |
| webEffects | Opt-in capabilities your device can contribute in collaborative spaces — for example, sharing your search index or acting as a backup host. Controlled from Sharing & Effects. |
Custom apps
| Term | Definition |
|---|---|
| Custom app | An app you install by uploading a .webai bundle from the launcher. It runs sandboxed with only the permissions declared in its manifest. See Build apps. |